关键词: 齿轮箱, 复杂工况, 故障诊断, 域对齐

Abstract: To address the insufficient cross-domain generalization capability of fault diagnosis models caused by significant distribution shifts between the source and target domains under complex and variable operating conditions, a gearbox fault diagnosis method based on cross-domain gradient collaborative optimization is proposed. The proposed approach first constructs a Multi-Scale Attention Aggregation (MSAA) network that fuses deep features of different scales across both channel and temporal dimensions,thereby enhancing the representation of domain-invariant features. On this basis,a bi-classifier framework is introduced,in which reliable pseudo-labels for the target domain are generated via softmax-weighted clustering to provide discriminative supervisory information for unlabeled samples. These pseudo-labels are then incorporated into the cross-domain gradient collaborative optimization mechanism,where the similarity of gradient directions between source and target domain samples at the shared classifier layer is used as a consistency constraint,effectively reducing inter-domain feature distribution discrepancies while preserving inter-class separability. Finally,the feature extractor and classifiers are jointly optimized through the integration of supervised loss,self-supervised loss,and gradient consistency loss,thereby significantly improving fault recognition performance in the target domain. Experimental results demonstrate that the proposed method achieves high diagnostic accuracy across various cross-condition tasks,validating its effectiveness and generalization capability under complex operating conditions.

Key words: gearbox, complex operating conditions, fault diagnosis, domain alignment